@companion-ai/feynman@0.3.45
Research-first CLI agent built on Pi and alphaXiv
AI Security Review
scanned 2mo ago · by lpm-firewall-aiInspectable install-time code is a Node version check, but the package still ships an unreadable 59MB runtime archive that the CLI extracts and a user-run installer that writes skills for other agents. That is dual-use payload and agent-extension risk without a concrete malicious install chain.
Decision evidence
public snapshot- The package ships a large opaque `.feynman/runtime-workspace.tgz` that cannot be read as text, then restores it on CLI start after a SHA-256 check.
- `bin/feynman.js` imports `scripts/patch-embedded-pi.mjs` on every launch, which extracts that archive into `.feynman/npm` and patches bundled Pi packages.
- A user-run skills installer copies downloaded prompts into Codex, OpenCode, or `.agents` skill directories.
- First-party tools spawn package managers and MCP connector processes; bundled vendor trees include HTTP credential providers and a WASM Doom example.
- Scanner AST coverage was incomplete and the tarball exceeds the inspectable size limit, so archive contents stay unresolved.
- npm `preinstall` only runs `scripts/check-node-version.mjs`, a Node version gate with no file, network, or agent-config mutation.
- There is no runtime self-dependency on `@companion-ai/feynman` and no unconsented preinstall write to a foreign agent control surface.
- Inspected first-party code is a research CLI wrapping Pi, with Hugging Face and science-database tools that match the stated product.
- `known_malware_source_similarity` hits are on bundled third-party files such as google-auth and highlight/marked, not a hidden dropper.
Source & flagged code
24 flaggedPackage defines install-time lifecycle scripts.
package.jsonView on unpkgPackage source references child process execution.
dist/setup/preview.jsView on unpkg · L1Findings
Affected versions and remediation
This report applies to @companion-ai/feynman@0.3.45.
See all versions of @companion-ai/feynman for other recorded verdicts.
- Review the evidence and your use of @companion-ai/feynman@0.3.45 before allowing it. Restrict the permissions described in this report.
- Choose an independently verified alternative or release. This report does not establish that other versions are safe.
Evidence last updated: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.